Face Swap Methods Based on Diffusion Models

Face swapping is a generative task where we replace the identity of a person in an image with a different one while preserving the semantic information such as the expression, pose, and background. One of the most common approaches in this field is the use of Generative Adversarial Networks. Recent works have utilized pre-trained models, like Style-GAN, in their methods. This approach has proven effective in achieving high-quality results in face swapping tasks. In this work, we will explore the capabilities of using modern diffusion-based approaches for face swapping tasks. We will build on the DiffFace method, which offers flexible inference and avoids the mode collapse problem. To improve the speed of inference, we will use the results from stable-diffusion. Our approach involves applying a pre-trained latent diffusion model and a continuous generation procedure. This allows for fast inference while maintaining high fidelity in the generated images.